HTCC poster for CERN Openlab opendays 2015

LHCb Upgrade

CERN openlab: www.cern.ch/openlab

High Throughput Computing Collaboration

Feasibility study to use Intel Xeon Phi and Xeon-FPGA computing accelerators with the Omni-Path network in the

Event Filter Farm for the LHCb Upgrade

Christian Färber, Karel Ha, Niko Neufeld, Rainer Schwemmer, Paolo Durante{christian.faerber, karel.ha, niko.neufeld, rainer.schwemmer, paolo.durante}@cern.ch

Physics Department, CERN, Geneva, Switzerland

Feasibility study to use Intel Xeon Phi and Xeon-FPGA computing accelerators with the Omni-Path network in the

Event Filter Farm for the LHCb Upgrade

Christian Färber, Karel Ha, Niko Neufeld, Rainer Schwemmer, Paolo Durante{christian.faerber, karel.ha, niko.neufeld, rainer.schwemmer, paolo.durante}@cern.ch

Physics Department, CERN, Geneva, Switzerland

Xeon Phi

After the Upgrade● 2018 LHCb will change its detector to a trigger-free

readout, reading every collision (one every 25 ns) and a much more flexible software-based trigger system, the Event Filter Farm (EFF).

● Events will be processed and triggered on an event-by-event basis by the Event Filter Farm.

The Event Filter Farm● Raw data input ~ 40 Tbit/s (already zero-suppressed by the front-end electronics)

● Only Software Trigger for selecting the events● EFF needs fast processing of trigger algorithms (decision within O(10) µs). → Different technologies have to be explored.

● High-speed interconnect technology has to be investigated and used.

512-bit vector units for SIMD parallelization

Support for C/C++, Fortran via OpenCL, OpenMP, Intel Cilk Plus

60+ x86 cores on KNC, even more on KNL

The current version is called Knights Corner (KNC),main focus put on the upcoming Knights Landing (KNL).

Xeon-FPGA

Interesting from the technical point of view● EFF needs bandwidth of 40 Tbit/s.● Many-to-many topology → Omni-Path support

Omni-Path

Built-in interconnect (Intel Omni-Path) on KNL

Coprocessor especially suitable for high throughput computing

Aims● Feasibility of Intel technology for LHCb Upgrade in an Online context

● Data acquisition and event-building: move from firmware to software

● Accelerator-assisted processing for high-level trigger

● Primarily, integer computations for event building: lookup tables, offset calculations etc.

(in contrast to float-intense operations for offline analysis and simulations)

● Cost vs. performance?

As stand-alone servers

Current situation ● Raw data output ~ 10 Tbit/s (not zero-suppressed)

● Hardware and Software Triggerfor selecting events

● Application example:Process ~ 1011 pp collisionsfor detecting one B

d

0 → µ+µ- ● Upgrade program foresees 10x higher pp collision rate.

Bd

0 → µ+µ-

● Intel's upcoming fabric technology suitable for DAQ systems● 100 Gbit/s● High-speed network across all Intel technology

→ “hand-tune” optimal ratio for HLT and Event Filter Farm (Xeon Phi vs. Xeon-FPGA vs. others)

16 GB of fast on-chip memory

CPU and FPGA are connected via the point-to-point interconnect QPI.

Attractive for Trigger algorithms● Track fitting● Pattern recognition● Particle identification

FPGA has cache-coherent memory access to the main memory of the server and

it can collaborate with the CPU.

First project: compareMuon Trigger performance on

CPU and on FPGA

There is a computing accelerator implemented on the FPGA,

which can be optimized for the algorithm.

FPGAs are already used in the current system in the first trigger stage. This makes a combined

system of CPU and FPGA very interesting for the future Event Filter Farm.

Test system is a two-socket platform from Intelwith a Xeon CPU and an FPGA.

And...● Code reviews● Training and consultancy for developers● Port LHCb code-base to icc compiler and provide maintenance

INTEL XEON PROCESSOR SERIES

FPGA

+